## Weekly signal

This was a practical infrastructure week for multi-agent systems. The strongest signal: enterprises are no longer treating agents as isolated copilots. The new focus is control planes, shared memory, agent identity, and operational handoffs. This briefing covers developments found through May 11, 2026; May 12, 2026 is still ahead of the current US date.

## What changed

1. Agent governance moved toward “fleet management.” Microsoft’s Agent 365 is now positioned as a control plane to observe, secure, and govern AI agents, with registry, activity mapping, Entra-based identity controls, Defender threat protection, and Purview data controls. ServiceNow pushed the same direction at Knowledge 2026, expanding AI Control Tower to discover, observe, govern, secure, and measure AI systems, agents, and workflows across clouds and enterprise apps. For multi-agent teams, this means the governance layer is becoming as important as the orchestration framework.

2. Shared memory became a product category. Yugabyte launched Meko, an agent-native data layer for multi-agent applications. Its core idea is a “datapack” that combines knowledge, memory, conversations, and traces, exposed through MCP so multiple agents can share learned context and preserve decision history. This directly targets a common production failure: Agent B receives a summary from Agent A but loses the assumptions, evidence, and reasoning behind it.

3. Security guidance shifted from prompt safety to agent identity and swarm behavior. OASIS/CoSAI published new work around Agentic Identity and Access Management and autonomous swarms after RSAC 2026. The key warning: as agents can spawn other agents, coordinate across systems, and act at machine speed, static access lists and pattern-based monitoring are not enough.

4. Open-source agent frameworks are adding durability, not just demos. NousResearch’s Hermes Agent v0.13.0 added “Multi-agent Kanban,” with durable boards, worker handoff, heartbeats, reclaim, retry budgets, zombie detection, and a hallucination gate. LangGraph4j’s May 7 release continued the Java path for stateful, multi-agent LLM apps with cycles, memory, collaboration, and handoffs.

5. Customer conversation stacks are becoming agent orchestration layers. Twilio announced generally available Conversation Memory, Conversation Orchestrator, Conversation Intelligence, and Agent Connect to manage persistent context across humans, AI agents, and systems.

## What to do with it

Audit every multi-agent prototype for four missing layers: agent registry, per-agent identity, shared memory, and traceable handoffs. If a system cannot answer “which agent acted, under whose authority, with what context, and why,” it is not production-ready. For builders, prioritize durable state, replayable traces, and least-privilege tool access before adding more specialized agents.